MediaTek
Artificial Intelligence
Leading the Edge-AI Technology Revolution
Edge AI
The rapid evolution of AI-enhanced use cases is changing how devices are being created and used. This shift in demand, as well as the increasing capability of AI technologies, is motivating products to process AI-enhancements ‘at the edge’ in products at home, in vehicle or about person, rather than rely solely on Cloud-connected support.
Advantages of this edge-computing include real-time immediacy, data privacy and overall lower power consumption. For device makers, there’s also no need to roll out supporting Cloud infrastructure into every region a product is sold into, decreasing time to market.
MediaTek Neural Processing Unit (NPU)
MediaTek develops its own Deep Learning Accelerators (Performance Cores), Visual Processing Units (Flexible Cores), hardware-based, multicore scheduler, and software development kits (NeuroPilot) that make up the core components of its industry-leading Neural Processing Units (NPUs).
MediaTek NeuroPilot
We’re meeting the Edge AI challenge head-on with MediaTek NeuroPilot. Through the heterogeneous computing capabilities in our SoC's such as NPUs, GPUs and CPUs, we are providing high-performance and power efficiency for AI features and applications. Developers can target these specific processing units within the chip, or, they can let MediaTek NeuroPilot SDK intelligently handle the processing allocation for them.
MediaTek Generative AI
As the industry leader in developing powerful, highly integrated and efficient system-on-chip products, MediaTek is enabling the future of AI by creating an ecosystem of edge-AI hardware processing paired with comprehensive software tools across its product range - smartphones and laptops to smart homes, intelligent IoT in businesses, and smart vehicles.
Today, tomorrow, and beyond
Today
As Artificial Intelligence (AI) continues to advance at a rapid pace, it’s reshaping the technology we use in our homes, workplaces and cities, bringing us new experiences and changing the way we interact. Today, AI enables technologies like deep-learning facial detection (DL-FD), real-time beautification with novel overlays, object and scene identification, AR/MR acceleration, real-time enhancements and augmentations to photography or video and much more.
The Future
The future of AI-enhanced devices is huge. Imagine using devices tailored perfectly to your needs and habits: a smartphone that keeps track of your health and orders you medicine before you get sick; a smart home that turns on the lights and heat just before you arrive; an autonomous car that drives you where you need to go, the moment you hop in. An innate intelligence that’s so fluid it brings a new level of user experience, and changes your world. That’s where Edge AI comes in.
The MediaTek NeuroPilot Advantage
Write Once, Apply Everywhere
MediaTek NeuroPilot SDK supports all MediaTek AI-capable hardware. It allows developers to ‘write once, apply everywhere’ for existing and future MediaTek hardware platforms and across all product lines, including smartphones, automotive, smart home, IoT and more. This streamlines the creation process, saving cost and time to market. The software ecosystem covers both Android and Linux OS’ and offers a complete compiler, profiler and application libraries.
Build-Friendly Frameworks
Applications can be built using common frameworks such as TensorFlow, TF Lite, Caffe, Caffe2 Amazon MXNet, Sony NNabla, or other custom 3rd party frameworks. At the API level for Android OS, Google Android Neural Networks API (Android NNAPI) and MediaTek NeuroPilot SDK are supported. The NeuroPilot SDK extends the Android NNAPI allowing developers and device makers to bring their code closer-to-metal for better performance and power-efficiency.
MediaTek Research
MediaTek Research commits itself to growing and elevating the AI ecosystem in everyday devices.
This includes sensors with extreme resolutions, sensors used exclusively for popular functions like bokeh capture, mono sensors for enhanced light sensitivity or specialist sensors for unique applications. Accompanying lenses can accommodate a multitude of capture styles including zoom, wide, macro, or just everyday photography.
Mar 8, 2023
Extending the Pre-Training of BLOOM for Improved Support of Traditional Chinese: Models, Methods and Results
In this paper we present the multilingual language model BLOOM-zh that features enhanced support for Traditional Chinese. BLOOM-zh has its origins in the open-source BLOOM models presented by BigScience in 2022.
Feb 02, 2023
Fisher-Legendre (FishLeg) optimization of deep neural networks
We introduce a new approach to estimate the natural gradient via Legendre-Fenchel duality, provide a convergence proof, and show competitive performance on a number of benchmarks.
Dec 19, 2022
A Learning-Based Algorithm for Early Floorplan With Flexible Blocks
This paper presents a learning-based algorithm using graph neural network (GNN) and deconvolution network to predict the placement of the locations and the aspect ratios for the design blocks with flexible rectangles.
Oct 31, 2022
Near-Optimal Collaborative Learning in Bandits
A near-optimal algorithm is proposed for pure exploration in a new framework for collaborative bandit learning that encompasses recent prior works.
Nov 24, 2022
Gradient Descent: Robustness to Adversarial Corruption
We provide performance guarantees for gradient descent under a general adversarial framework
Improved Convergence Rates for Sparse Approximation Methods in Kernel-Based Learning
Kernel-based models such as kernel ridge regression and Gaussian processes are ubiquitous in machine learning applications for regression and optimization.
Feb 20, 2022
Regret Bounds for Noise-Free Kernel-Based Bandits
Kernel-based bandit is an extensively studied black-box optimization problem, in which the objective function is assumed to live in a known reproducing kernel Hilbert space.
LPI: Learned Positional Invariances for Transfer of Task Structure and Zero-shot Planning
Real-world tasks often include interactions with the environment where our actions can drastically change the available or desirable long-term outcomes.
Jun 1, 2022
Adaptive erasure of spurious sequences in sensory cortical circuits
Sequential activity reflecting previously experienced temporal sequences is considered a hallmark of learning across cortical areas.
Apr 13, 2022
Flexible Multiple-Objective Reinforcement Learning for Chip Placement
Recently, successful applications of reinforcement learning to chip placement have emerged. Pretrained models are necessary to improve efficiency and effectiveness.
